Adarve-Castro Antonio, Soria-Utrilla Virginia, Castro-García José Miguel, Domínguez-Pinos María Dolores, Sendra-Portero Francisco, Ruiz-Gómez Miguel J, Algarra-García José
Departamento de Radiología y Medicina Física, Facultad de Medicina, Universidad de Málaga, Bulevar Louis Pasteur, 32, 29010, Málaga, Spain.
University Hospital Virgen de la Victoria, Málaga, Spain.
Radiol Med. 2025 Apr 24. doi: 10.1007/s11547-025-02004-z.
This study aims to assess the proficiency of supervised machine learning techniques in discriminating between normal and abnormal bone mineral density (BMD) by leveraging clinical features and texture analysis of spinal bone tissue in patients diagnosed with primary hyperparathyroidism (PHP). From a total of 219 patients diagnosed with PHP, the 58 who had undergone both DXA and abdominal CT scan were included in this study. BMD was assessed by quantifying the Hounsfield units (HU) and performing texture analysis on every CT scan. The first lumbar vertebral body texture features were extracted by using LifeX 7.3.0 software. Initial classification into normal or abnormal BMD was performed with different machine learning techniques by training a model with the variables obtained from the texture analysis. Differentiating osteopenia from osteoporosis was evaluated by creating two models, one including the variables obtained from the texture analysis and HU and another one which only included the HU. Their performance was evaluated in the validation and test groups by calculating the accuracy, precision, recall, F1 score, and AUC. Bayes demonstrated higher performance for discerning individuals with normal and abnormal BMD, with an AUC of 0.916. The results from the second analysis showed a better performance for the model including the variables obtained from the texture analysis compared to the model that was solely trained with the HU (AUC in the training group of 0.77 vs. 0.65 in the test groups, respectively). In conclusion, analysis of BMD obtained from abdominal CT scans including texture analysis provide a better classification of normal density, osteopenia and osteoporosis in patients with PHP.
本研究旨在通过利用原发性甲状旁腺功能亢进症（PHP）患者脊柱骨组织的临床特征和纹理分析，评估监督式机器学习技术在区分正常和异常骨密度（BMD）方面的熟练程度。在总共219例诊断为PHP的患者中，本研究纳入了58例同时接受双能X线吸收法（DXA）和腹部CT扫描的患者。通过量化亨氏单位（HU）并对每次CT扫描进行纹理分析来评估骨密度。使用LifeX 7.3.0软件提取第一腰椎椎体的纹理特征。通过使用从纹理分析获得的变量训练模型，用不同的机器学习技术对骨密度进行正常或异常的初步分类。通过创建两个模型来评估区分骨质减少和骨质疏松症，一个模型包括从纹理分析和HU获得的变量，另一个模型仅包括HU。通过计算准确率、精确率、召回率、F1分数和曲线下面积（AUC），在验证组和测试组中评估它们的性能。贝叶斯方法在辨别骨密度正常和异常个体方面表现出更高的性能，AUC为0.916。第二次分析的结果表明，与仅用HU训练的模型相比，包括从纹理分析获得的变量的模型表现更好（训练组的AUC分别为0.77，测试组为0.65）。总之，对腹部CT扫描获得的骨密度进行分析，包括纹理分析，可为PHP患者的正常密度、骨质减少和骨质疏松症提供更好的分类。
